Motivation
A linear equation can be specified without writing an operator matrix. Instead, require an unknown vector to satisfy against every test vector . A bounded form records how the unknown interacts with each test. Coercivity says that testing a vector against itself controls its norm from below.
That lower bound does more than prove uniqueness. Combined with Hilbert-space orthogonality, it also supplies existence for every bounded right-hand side. Restricting the test and trial vectors to a smaller closed subspace then gives an approximation problem with the same structure. The resulting error estimate explains when increasing finite-dimensional systems approximate the original infinite-dimensional solution.
We will distinguish three claims: solving a variational equation, minimizing an energy, and approximating its solution. They coincide in an important symmetric setting but are not interchangeable in general. The proof uses Riesz representation, completeness and the Hilbert projection theorem; no compactness theorem is needed.
Definitions
Definition 2.7.1
(Bounded coercive forms and the complex convention)
Let be a real or complex Hilbert space with inner product linear in its first argument. In the real case, is bilinear. In the complex case it is linear in the first variable and conjugate-linear in the second. Assume constants satisfy
The first condition is boundedness; the second is coercivity. In the real case the real-part symbol can be omitted. Hermitian symmetry is the separate condition , reducing to symmetry over the reals.
The two constants have different jobs. Boundedness controls interactions of different vectors; coercivity supplies a uniform positive diagonal estimate. A form may be coercive without being Hermitian. Conversely, invertibility alone need not supply a positive real-part estimate.
Definition 2.7.2
(Right-hand sides and the variational problem)
Because is conjugate-linear in the complex case, the right-hand side must be a bounded conjugate-linear functional. Its norm is . Riesz representation gives a unique with
The problem is to find such that for every . Over the reals, conjugate-linearity is simply linearity.
This convention does not redefine the usual linear dual. If one starts with a bounded linear functional on a complex space, then has the appropriate placement here. The stated representation follows from ordinary Riesz applied to : conjugate its representing identity. Using a linear right-hand side against a conjugate-linear test slot would make the equations inconsistent under replacing by , except in the zero case.
Theorem/Proposition
Theorem 2.7.3
(Lax–Milgram: existence, uniqueness and stability)
Under the boundedness and coercivity assumptions, every bounded right-hand side of the stated type has a unique solution. It satisfies
For two right-hand sides, . No symmetry assumption on is required.
Theorem 2.7.4
(The energy interpretation requires Hermitian symmetry)
If in addition is Hermitian, then is real and
has the unique global minimizer supplied by Lax–Milgram. For every ,
Conversely, a minimizer of this Hermitian energy solves the variational equation. The same scalar quadratic formula need not represent a nonsymmetric variational problem.
Theorem 2.7.5
(Galerkin approximation and the Céa estimate)
Let be a closed linear subspace. There is a unique such that for every . If is the full solution, then
Thus for nested closed subspaces with dense union, in norm. Finite-dimensional subspaces are closed, so the theorem applies in particular to finite Galerkin systems.
The equation called Galerkin orthogonality is orthogonality with respect to the form, not automatically with respect to the original Hilbert inner product. In the Hermitian case, defines its own energy inner product, and the approximation is the exact best approximation in that energy norm. The general estimate above compares instead in the original norm, with factor .
Proof sketch or proof idea
Proof
(Represent the form, then prove the range is closed and dense)
For each fixed , Riesz representation of gives a vector with . Uniqueness of representation makes linear, and boundedness gives . Coercivity and Cauchy–Schwarz imply
Hence is injective. Its range is closed: if converges, the lower bound makes Cauchy, completeness gives , and continuity gives the output limit .
The range is also dense. If is perpendicular to it, then for every . In particular , so coercivity forces . The Hilbert projection theorem now shows that the closure of the range is all of . Closedness and density together give surjectivity.
Represent by , and solve . The lower bound gives the claimed estimate. Subtracting the equations for two right-hand sides proves the stability estimate. This completes the existence proof rather than assuming that a lower bound alone implies surjectivity.
Proof X-ray
Two different uses of coercivity
The lower norm bound supplies injectivity and closed range. It does not by itself prove that every target lies in the range. The second step tests a vector perpendicular to the range against itself; the positive real-part estimate makes that vector zero. Hilbert orthogonality supplies density, and only then does closed range become the whole space.
For the Hermitian energy, expand . Symmetry makes the cross terms sum to , which cancels because solves the equation. This proves the exact gap. Conversely, differentiating at real gives . In a complex space also use the direction to get its imaginary part zero. Thus the full equation follows.
Proof
(Restrict to a closed subspace and compare the error)
The closed subspace is itself Hilbert. Restricted boundedness and coercivity have the same constants, so Lax–Milgram on supplies . Subtraction gives for and every .
For any , decompose the second argument . Additivity and the vanishing second term give
If , divide by its norm and take the infimum; if , the estimate is immediate. Nestedness and density make the infimum tend to zero, proving convergence.
When is Hermitian, expand the energy norm of . The cross term is zero by Galerkin orthogonality. Therefore , proving exact best approximation in that norm.
Proof X-ray
Why the constant is the ratio of the two form constants
Closedness of is used once: it makes a Hilbert space on which the same and hold, so Lax–Milgram applies there. The real idea is Galerkin orthogonality. Because , it lets the second argument of be replaced by for any competitor at no cost. Coercivity then bounds from below by , and boundedness bounds from above by ; the quotient is exactly what remains. Neither dimension nor symmetry is used.
For a finite basis of , write . The coefficient matrix has entries , and the right-hand side is . The order matters for a nonsymmetric form. Coercivity makes this finite system nonsingular: a homogeneous solution would have for every , including , forcing . A poorly scaled basis can still affect numerical conditioning; the abstract error bound is about vectors in , not arbitrary coordinate norms.
Worked examples
Worked example 2.7.6
(A diagonal form and a controlled truncation)
On , let
The series is absolutely convergent by Cauchy–Schwarz. One may take and . The solution is , and .
For , the Galerkin solution keeps exactly the first coordinates of . It is even the ordinary orthogonal projection in this particular diagonal setting. If , then and
The explicit tail gives a sharper conclusion than simply knowing that a dense family will eventually approximate the solution.
Worked example 2.7.7
(A weak boundary-value equation and a one-dimensional trial space)
Let consist of absolutely continuous functions on with zero endpoint values and derivative in . Use . This is Hilbert: the derivative map is an isometry onto the closed subspace , with inverse .
Also , by Cauchy–Schwarz applied to that integral. Thus is bounded. With , Lax–Milgram gives a unique weak solution. The explicit function satisfies the identity by integration by parts, since and has zero endpoints. No general regularity theorem is being used to infer smoothness of every weak solution here.
Take with . Since and , the approximation is . Direct calculation gives
Here the form is the Hilbert inner product itself, so the discrete solution is the best approximation in the derivative norm.
Worked example 2.7.8
(Coercive and nonsymmetric is not the same as energy minimization)
On , take and . Its skew part disappears from , so ; since , works. With , the variational solution is .
But the scalar expression is minimized at . It has forgotten the skew part of . Lax–Milgram solves the original nonsymmetric equation correctly; the Hermitian energy theorem cannot be applied to this form. A positive diagonal estimate is not a license to attach the wrong minimization problem.
Common mistakes
Counterexample mode
Bounded below does not mean onto
The right shift on is an isometry, so it is bounded below by one and has closed range. It is not onto: no image has first coordinate one. The associated form is not coercive, since . This does not contradict Lax–Milgram. It identifies why its dense-range step needs more than the output norm estimate.
The positive real-part estimate is a hypothesis of a sufficient condition. Two misreadings of it recur.
Common mistake
(Treating coercivity as necessary for invertibility)
Coercivity is sufficient, not necessary. On a nonzero complex Hilbert space, has for every , yet its operator is invertible and every variational equation has the unique solution . Failure of coercivity removes the theorem's guarantee; it does not prove non-solvability.
Common mistake
(Expecting convergence from a dense union alone)
The Céa estimate controls by the distance from to . Convergence of a Galerkin sequence therefore requires . Nested subspaces with dense union guarantee this; a non-nested family with dense union need not, as the exercise below on non-nested subspaces shows.
Summary
Bounded coercive forms define invertible Hilbert-space operators through Riesz representation. Coercivity supplies both a quantitative lower bound and the missing dense-range argument. Hermitian symmetry adds an exact energy interpretation. Restricting the variational problem to closed subspaces yields Galerkin orthogonality and a best-approximation bound, while the examples show why symmetry, test-slot convention and approximation-space hypotheses must remain visible.
Exercises
Quick check
(On a nonzero complex Hilbert space, let and . When is the stated real-part coercivity available? Solve the equation and compare the actual norm of the solution with the theorem's bound.)
Also explain why the right-hand side has the correct conjugate-linearity.
Solution · Answer
The real diagonal part is , so positive coercivity is available exactly when , with and . The solution is , hence ; the theorem gives the potentially weaker bound . At , invertibility remains even though this coercivity condition fails. The functional is conjugate-linear in , matching . Using in that slot would change its scalar behavior and generally would not define the same solvable equation.
Quick check
(On , let with , and . Find the full solution and the Galerkin solution in , compare with the Euclidean orthogonal projection, and compute the squared energy-norm error.)
Do not assume that the form's orthogonality is Euclidean orthogonality.
Solution · Answer
The full solution is . Write ; testing with gives , so . The Euclidean projection of is , a different point. The error is and
The Galerkin vector minimizes error in the energy norm defined by , not in the unweighted Euclidean norm.
Quick check
(For the inner-product form on with right-hand vector , set and for . Show that the union is dense but the Galerkin sequence does not converge to .)
Identify the missing approximation hypothesis rather than disputing the error estimate.
Solution · Answer
Every coordinate vector eventually belongs to an odd-indexed space, so the union contains the finitely supported sequences and is dense. For this form the Galerkin solution is the orthogonal projection of . At every odd stage the projection of is , while at every even stage it is zero. The sequence therefore does not converge. Its best-approximation distances alternate between zero and one, so the error estimate predicts no convergence. The spaces are not nested; the theorem's sufficient convergence assumption has not been met.